[PATCH] D134456: [PGO] Consider parent context when weighing branches with likelyhood.

[Aaron Ballman via Phabricator via cfe-commits]

aaron.ballman added a comment.

In D134456#3818982 <https://reviews.llvm.org/D134456#3818982>, @rnk wrote:

> In D134456#3818952 <https://reviews.llvm.org/D134456#3818952>, @aaron.ballman wrote:
>
>> I think that would be reasonable to at least consider; the default behavior can be "the attribute wins" and this gives users an escape hatch to say "no, I trust PGO more, let that win instead".
>
> Right, but that's a policy change. We'd need to update docs, update release notes, and all of that. We'll need this fix in any case for the mode where PGO wins, which will remain long term for the users that rely on it.
>
> What commitments can we make to land this fix without making a policy change, which is quite a great deal of work (RFC, docs, notes, etc)?

Personally, I consider it a bug fix and not a policy change (which doesn't mean we don't need to make all these updates, of course). We didn't carefully consider this aspect of the original design when doing the implementation work, now we'd like to correct that as responsibly as we can.

Alternatively, we could document that we don't follow the intent of the standard feature and that's just the way things are, and add a command line option to honor that intent. However, given that GCC honors the attribute over PGO, I think we should change our default behavior.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D134456/new/

https://reviews.llvm.org/D134456